Dialectica AI is a public portfolio demonstration, not a hardened multi-tenant production service.

• Never commit .env or .streamlit/secrets.toml.
• Use a dedicated OpenAI project key for the Streamlit deployment.
• Store OPENAI_API_KEY only in local environment configuration or Streamlit Secrets.
• Store the same long random MCP_AUTH_TOKEN in Streamlit Secrets and Render environment variables.
• Rotate both credentials immediately if either appears in source, logs, screenshots, or issue reports.
• Use restricted API-key permissions where available.

• /healthz is intentionally unauthenticated and exposes only non-sensitive service statistics.
• /mcp requires bearer authentication in cloud mode.
• The public MCP contract excludes uploads, arbitrary indexing, URL fetching, and index-reset tools.
• The prepared corpus uses local deterministic embeddings, so the MCP service does not need an OpenAI key.
• The application does not send complete papers to OpenAI; model calls receive retrieved curator-written notes and metadata.

• The app enforces soft browser-session analysis limits.
• These limits can be bypassed by a determined visitor and must not be treated as a hard spending cap.
• Use a dedicated OpenAI project, restricted credentials, usage alerts, conservative limits, and regular usage review.
• Avoid publishing the application widely until usage and failure behaviour have been tested.

• Indexed text is curator-written summary material, not direct paper quotations.
• Every note retains attribution and a link to the source paper.
• The demo should not be presented as a systematic literature review or as an authoritative scientific consensus engine.

Do not place API keys, bearer tokens, private documents, or full deployment logs in a public issue. Revoke exposed credentials before sharing a sanitized report.